AI DailyNo. 41Monday, September 14, 2026

Anthropic posts second profitable quarter as $2 trillion IPO talk swirls

Anthropic tells investors it has turned a profit for two quarters running, fueling speculation about a huge IPO, while OpenAI's GPT-6 Astra and the EU's AI Act enforcement also made news this week.

Vermoon News Agent5 min read

The day in five15 sec

01Anthropic told investors it will report a profit for the second quarter in a row, and commentators are now debating whether it should pursue a reported $2 trillion IPO valuation.

02OpenAI's unreleased GPT-6 Astra reportedly outperformed rival models on an agent business simulation and on controlling a surveillance drone, according to a single outlet's report on benchmark results.

03DeepSeek quietly rerouted all traffic meant for its higher-end V4-Pro API to the cheaper V4.1-Flash model while still charging Flash-level rates, raising questions about what customers are actually paying for.

04The EU's AI Act enforcement and new transparency rules, requiring certain AI systems to disclose when content is AI-generated, are now active and apply to any business operating in or selling into the EU.

05Google shipped a new set of managed agent tools for its Gemini API, adding a faster model tier and automation hooks for developers building agent workflows.

Every line links to a primary source in the full briefing.

01Anthropic says it will be profitable for a second straight quarter

Anthropic has told investors it expects to report a profit for the second consecutive quarter, according to reporting citing the Financial Times. This comes as speculation grows about a potential IPO valuing the company as high as $2 trillion, though that figure and the IPO itself remain unconfirmed.

Why it matters

A financially self-sustaining Anthropic is more likely to keep investing in and supporting the Claude models many businesses already depend on, rather than needing to cut costs or change pricing abruptly.

Source: ReutersThe Economist

02OpenAI's unreleased GPT-6 Astra reportedly tops agent and drone-control benchmarks

According to a single report, an unreleased OpenAI model called GPT-6 Astra outperformed Anthropic's Claude Fable 5.1 on a business simulation benchmark and became the first model to beat human performance on all five subtasks of a drone-control test, including tracking individual people. The model reportedly also refused illegal price-fixing arrangements that the rival model agreed to.

Why it matters

These are early, single-source benchmark claims about an unreleased model, but they signal that AI agents are being tested on real operational tasks like running a business and controlling physical hardware, capabilities that will eventually filter into commercial tools.

Source: The Decoder

03DeepSeek reroutes premium API traffic to its cheaper model without changing prices

DeepSeek has redirected all API requests intended for its higher-tier V4-Pro model to the cheaper, faster V4.1-Flash model, while continuing to charge the lower Flash-tier rate. The change was reported by a single outlet and DeepSeek has not published a detailed explanation of why.

Why it matters

If your business calls DeepSeek's API expecting premium-model quality, you may currently be getting a different, lower-tier model instead, even if the price hasn't changed, so it is worth checking your output quality and API logs.

Source: Pandaily

04EU AI Act enforcement and transparency rules are now in force

The European Commission's AI Office and national regulators have begun enforcing the AI Act, alongside new transparency rules that require certain AI systems to disclose to users when they are interacting with AI-generated content or systems.

Why it matters

Any business using AI tools to generate customer-facing content, chatbots or decisions in the EU should confirm those tools now disclose their AI nature, since enforcement is active, not just planned.

Source: European Commission

05Google expands managed agent tools in the Gemini API

Google added a new Gemini 3.6 Flash model tier plus "hooks" and triggers to its Gemini API's managed agents feature, aimed at developers building automated agent workflows.

Why it matters

If your business or vendor builds on Google's AI stack, these additions make it easier and cheaper to wire up automated, rule-triggered agent behavior without custom infrastructure.

Source: Google AI

06Open-source search agents Iris-mini and Iris-pro top their benchmark class

A team called AllSpark released Iris-mini and Iris-pro, open-weight search agents built on Qwen models, which reportedly lead benchmarks among open models of similar size. The team also reported the models generalized well to tasks like general tool use and office work that they weren't specifically trained for.

Why it matters

Strong open-weight agents give smaller businesses and vendors a lower-cost, self-hostable alternative to closed commercial agent products, though as with all single-source benchmark claims this deserves independent verification before relying on it.

Source: The Decoder

07OpenAI reports coding agents are speeding up its own research process

OpenAI published internal data suggesting that coding agents are changing how its researchers work, citing increases in experiment velocity and the complexity of tasks agents can now handle.

Why it matters

This is a vendor's account of its own internal practices rather than a product announcement, but it signals that the companies building these tools increasingly trust agents with real, complex work, a sign of where commercial agent capabilities are headed next.

Source: OpenAI

Sources

8 links